Skip to content

TASKS.md

Latest commit

 

History

History
145 lines (122 loc) · 7.68 KB

TASKS.md

File metadata and controls

145 lines (122 loc) · 7.68 KB
Raw

Tasks

2026-09-24: Idiot-proof DATEV export (DONE)

Why: A user report showed the DATEV feature was unusable: the exporter ignored the configured consultant/client numbers, wrote no EXTF header (DATEV refuses such files), used a simplified column layout, emitted UTF-8 instead of CP1252, silently booked unknown tax rates as 19%, and the documented CLI command did not exist.

What shipped:

  • src/datev/ core: format.ts (EXTF header v700, Buchungsstapel format version 13 with all 125 columns, CP1252 encoding), config.ts (zod-validated settings stored in data/config.json, DATEV_* env overrides), accounts.ts (SKR03/SKR04 with per-category overrides), exporter.ts (strict batch builder with visible warnings), export-run.ts (skip-and-report per broken receipt, automatic one-file-per-fiscal-year split), receipt-cache.ts (SQLite cache so re-runs are instant), service.ts (tag+date selection, cache-first extraction).
  • src/cli/datev.ts: npm run datev:init (interactive wizard, validates every answer, works with piped input), datev:check (dry run with full report), datev:export.
  • MCP tool papercortex_export writes real files and reports skips; missing configuration returns instructions instead of exporting garbage.
  • docs/datev.md step-by-step guide (German, English summary), README and docs updated, eslint.config.js added (lint previously crashed on every machine: ESLint 9 without any config).

Evidence: 40/40 vitest, tsc --noEmit clean, lint green, build green, wizard smoke-tested (piped answers + EOF abort), unconfigured datev:check prints the init pointer.

Open: none for this feature. GitHub issues #1 (compose restart loop) and #2 (initial vector store population) remain separate onboarding tasks.

2026-09-24: Web UI settings page for DATEV (DONE)

Why: Maintainer: the DATEV setup must be idiot-proof for Docker/UnRaid users -- "eventuell langt eine Einstellungsseite". The CLI wizard covers terminal users; the target audience lives in browser UIs.

What shipped:

  • src/webui/: api.ts (settings round-trip, system check, preview, export -- dependency-injected and unit-tested), jobs.ts (background extraction with progress polling), server.ts (routing, optional WEBUI_TOKEN auth, CSP, strict export-filename validation), page.ts (self-contained German single-page UI: Systemcheck / DATEV-Einstellungen / Belege exportieren), start.ts + main.ts (embedded in the MCP server process by default, standalone via npm run webui).
  • System check with actionable diagnostics ("ollama pull ", "PAPERLESS_URL prüfen"); settings validated field-by-field with German error messages; export flow: load -> analyze (progress bar) -> preview with warnings and per-receipt opt-out -> CP1252 EXTF download.
  • Robustness found by smoke test: better-sqlite3 dies on a fresh checkout because ./data does not exist -- vector store and receipt cache now create their directory (very likely the mechanism behind GitHub issue #1).
  • docker-compose port 8140 (localhost-bound by default), Dockerfile EXPOSE, .env.example WEBUI_* block, docs/datev.md + README sections.

Evidence: 51/51 vitest (16 new webui tests incl. path-traversal guard on downloads), tsc, lint, build green; UI verified in the browser end-to-end (doctor renders, invalid Beraternummer rejected with field error, valid save persists config.json, banner confirms).

Coordination: built on top of the consolidated src/datev core from the parallel session (see entry above); no core files changed except the mkdir robustness fix. Push and production deploy pending maintainer review.

2026-09-24: Web UI security hardening (DONE)

Why: Automated security review on the web UI commit: access code leaked into download URLs, port open to the LAN by default without a code, no DNS-rebinding defense, fail-open when no code is set.

What shipped:

  • Download tickets (src/webui/tickets.ts): single-use, 60s TTL, bounded; the access code lives only in the X-Auth-Token header and never in a URL.
  • Host-header allowlist (localhost by default, WEBUI_ALLOWED_HOSTS for LAN names/IPs) as the DNS-rebinding defense; 403 carries the exact fix.
  • Constant-time access-code comparison; loud startup warning when no code is configured.
  • Shipped compose now binds 127.0.0.1:8140 with a documented LAN variant; docs and .env.example explain the tradeoff.

Evidence: 59/59 vitest incl. real-HTTP server tests (foreign Host rejected with hint, 401 paths, ticket single-use and replay refusal, path-traversal ticket refusal), tsc, lint, build green.

2026-09-24: Extraction robustness, sanity checks, index builder (DONE)

Why: Maintainer request: work through the prioritized improvement list. The extractor did a bare JSON.parse on LLM output (breaks on markdown fences), amounts were never cross-checked, twice-scanned receipts booked twice, and NOTHING ever populated the vector store (GitHub issue #2 -- semantic search returned nothing forever).

What shipped:

  • src/receipt/extraction-parse.ts: fence-stripping, string-aware balanced-brace JSON extraction, zod-validated result with documented per-field defaults (unparseable date stays as-is so the exporter SKIPS with the real reason instead of booking "today"); extractor retries once with a sterner instruction, then fails with a clear error.
  • Exporter sanity checks: net+tax vs gross and tax-amount vs tax-rate arithmetic warnings; duplicate detection (vendor+date+amount) flagged in the preview so the user unticks the second scan BEFORE export.
  • src/webui/indexer.ts + "Suchindex aufbauen" button in the system check: paginated full index build as a background job with progress, skip already-indexed, per-document errors -- closes issue #2.
  • .github/workflows/docker-publish.yml: multi-arch ghcr.io image on push to main and v* tags (activates with the pending push).

Evidence: 77/77 vitest (18 new: extraction-parse 6, extractor retry 4, indexer 3, exporter sanity/duplicates 4, adjusted fakes), tsc, lint, build green.

Open: UnRaid Community-Apps template once the first ghcr image exists; vision-model extraction and payment-method money accounts remain backlog.

2026-09-24: Backlog round -- vision, payment accounts, matching, lock, i18n (DONE)

Why: Maintainer request: clear the parked backlog, plus bilingual support (de/en).

What shipped:

  • Vision extraction: optional OLLAMA_VISION_MODEL; when OCR is thin (<150 chars) or the text path fails, the model reads the receipt image (original file for images, rendered thumbnail for PDFs). Doctor checks the model when configured. No model set = unchanged behavior.
  • Money account per payment kind: cash books against the Kasse (1000/1600 default), card optionally against its own account; free-text methods normalized (bar/EC/girocard/...). Preview shows Zahlart -> Gegenkonto per booking.
  • Festschreibung toggle (GoBD): header field 21 and row field 114.
  • Bank reconciliation wired up (was dead code): parseBankCsvText for uploads, German date normalization DD.MM.YYYY (previously Invalid Date made the date signal silently useless), vendor matching on the first meaningful word instead of a fixed 8-char prefix ("REWE SAGT DANKE"), POST /api/match with 1MB CSV limit, step 4 in the export tab: matched / receipts-without-transaction / open transactions. Read-only.
  • Bilingual UI (de/en): dictionary-based, toggle persists, doctor labels localized. Booking warnings and the DATEV report stay German by design (they address the German tax advisor and travel into the export files).

Evidence: 87/87 vitest (10 new), tsc, lint, build green; UI smoke in the browser in both languages incl. the live vision-model doctor check.